Understanding cell fate decisions with temporal attention.

A Transformer model predicts cancer cell fate from raw video sequences with 94% accuracy, identifying predictive signals up to 8 hours before cell outcomes.

• Why it matters: Understanding non-genetic factors influencing cell fate is vital for improving cancer therapies, as genetically identical cells can respond differently under the same treatment.
• What they did: A deep learning approach using a Transformer model was trained on long-term live-cell recordings to predict cell fate directly from image sequences, without predefined features, and includes an explainability framework.
• The result: The model's predictions reveal that key signals are distributed throughout the cell trajectory, with insights into the roles of cell morphology, p53 signaling, and temporal cues in cellular decision-making.